GenCast, a new AI model from Google DeepMind, is accurate enough to compete with traditional weather forecasting. It managed to outperform a leading forecast model when tested on data from 2019, according to recently published research.

AI isn’t going to replace traditional forecasting anytime soon, but it could add to the arsenal of tools used to predict the weather and warn the public about severe storms. GenCast is one of several AI weather forecasting models being developed that might lead to more accurate forecasts.

“Weather basically touches every aspect of our lives … it’s also one of the big scientific challenges, predicting the weather,” says Ilan Price, a senior research scientist at DeepMind. “Google DeepMind has a mission to advance AI for the benefit of humanity. And I think this is one important way, one important contribution on that front.”

Price and his colleagues tested GenCast against the ENS system, one of the world’s top-tier models for forecasting that’s run by the European Centre for Medium-Range Weather Forecasts (ECMWF). GenCast outperformed ENS 97.2 percent of the time, according to research published this week in the journal Nature.

GenCast is a machine learning weather prediction model trained on weather data from 1979 to 2018. The model learns to recognize patterns in the four decades of historical data and uses that to make predictions about what might happen in the future. That’s very different from how traditional models like ENS work, which still rely on supercomputers to solve complex equations in order to simulate the physics of the atmosphere. Both GenCast and ENS produce ensemble forecasts, which offer a range of possible scenarios.

When it comes to predicting the path of a tropical cyclone, for example, GenCast was able to give an additional 12 hours of advance warning on average. GenCast was generally better at predicting cyclone tracks, extreme weather, and wind power production up to 15 days in advance.

One caveat is that GenCast tested itself against an older version of ENS, which now operates at a higher resolution. The peer-reviewed research compares GenCast predictions to ENS forecasts for 2019, seeing how close each model got to real-world conditions that year. The ENS system has improved significantly since 2019, according to ECMWF machine learning coordinator Matt Chantry. That makes it difficult to say how well GenCast might perform against ENS today.

To be sure, resolution isn’t the only important factor when it comes to making strong predictions. ENS was already working at a slightly higher resolution than GenCast in 2019, and GenCast still managed to beat it. DeepMind says it conducted similar studies on data from 2020 to 2022 and found similar results, although that hasn’t been peer-reviewed. But it didn’t have the data to make comparisons for 2023, when ENS started running at a significantly higher resolution.

Dividing the world into a grid, GenCast operates at 0.25 degree resolution — meaning each square on that grid is a quarter degree latitude by quarter degree longitude. ENS, in comparison, used 0.2 degree resolution in 2019 and is at 0.1 degree resolution now.

Nevertheless, the development of GenCast “marks a significant milestone in the evolution of weather forecasting,” Chantry said in an emailed statement. Alongside ENS, the ECMWF says it’s also running its own version of a machine learning system. Chantry says it “takes some inspiration from GenCast.”

Speed is an advantage for GenCast. It can produce one 15-day forecast in just eight minutes using a single Google Cloud TPU v5. Physics-based models like ENS might need several hours to do the same thing. GenCast bypasses all the equations ENS has to solve, which is why it takes less time and computational power to produce a forecast.

“Computationally, it’s orders of magnitude more expensive to run traditional forecasts compared to a model like Gencast,” Price says.

That efficiency might ease some of the concerns about the environmental impact of energy-hungry AI data centers, which have already contributed to Google’s greenhouse gas emissions climbing in recent years. But it’s hard to suss out how GenCast compares to physics-based models when it comes to sustainability without knowing how much energy is used to train the machine learning model.

There are still improvements GenCast can make, including potentially scaling up to a higher resolution. Moreover, GenCast puts out predictions at 12-hour intervals compared to traditional models that typically do so in shorter intervals. That can make a difference for how these forecasts can be used in the real world (to assess how much wind power will be available, for instance).

“You would want to know what the wind is going to be doing throughout the day, not just at 6AM and 6PM,” says Stephen Mullens, an assistant instructional professor of meteorology at the University of Florida who was not involved in the GenCast research.

While there’s growing interest in how AI can be used to improve forecasts, it still has to prove itself. “People are looking at it. I don’t think that the meteorological community as a whole is bought and sold on it,” Mullens says. “We are trained scientists who think in terms of physics … and because AI fundamentally isn’t that, then there’s still an element where we’re kind of wrapping our heads around, is this good? And why?”

Forecasters can check out GenCast for themselves; DeepMind released the code for its open-source model. Price says he sees GenCast and more improved AI models being used in the real world alongside traditional models. “Once these models get into the hands of practitioners, it further builds trust and confidence,” Price says. “We really want this to have a kind of widespread social impact.”